Constraint defined DEFERRABLE INITIALLY IMMEDIATE is still DEFERRED?

There may be a slight documentation bug here, but not for the case you're showing. If you BEGIN a transaction and try the updates one at time, they fail, but if a single statement leaves things in a good state, it doesn't complain. The docs say:

If a constraint is deferrable, this clause specifies the default time to check the constraint. If the constraint is INITIALLY IMMEDIATE, it is checked after each statement. This is the default. If the constraint is INITIALLY DEFERRED, it is checked only at the end of the transaction.

Which is exactly what seems to be happening. What is the surprise to me, given the documentation of DEFERRABLE, which says in part:

A constraint that is not deferrable will be checked immediately after every command.

Without the DEFERRABLE INITIALLY IMMEDIATE options, the example update fails, even though the UPDATE statement (presumably constituting the "command") leaves things in a good state. Perhaps the docs should be modified to say that a NOT DEFERRABLE constraint is enforced as each row is modified by a statement?

I remember having raised an almost identical point when PG9 was in alpha state. Here was the answer from Tom Lane (high-profile PG core developer):
http://archives.postgresql.org/pgsql-general/2010-01/msg00221.php

In short: won't fix.

Not to say that I agree with your suggestion that the current behavior is a bug. Look at it from the opposite angle: it's the behavior of NOT DEFERRABLE that is incorrect.

In fact, the constraint violation in this UPDATE should never happen in any case, since at the end of the UPDATE the constraint is satisfied. The state at the end of the command is what matters. The intermediate states during the execution of a single statement should not be exposed to the user.

It seems like the PostgreSQL implements the non deferrable constraint by checking for duplicates after every row updated and failing immediately upon the first duplicate, which is essentially flawed. But this is a known problem, probably as old as PostgreSQL. Nowadays the workaround for this is precisely to use a DEFERRABLE constraint. And there is some irony in that you're looking at it as deficient because it fails to fail, while somehow it's supposed to be the solution to the failure in the first place!

Summary of the status quo since PostgreSQL 9.1

• NOT DEFERRABLE UNIQUE or PRIMARY KEY constraints are checked after each row.

• DEFERRABLE constraints set to IMMEDIATE (INITIALLY IMMEDIATE or via SET CONSTRAINTS) are checked after each statement.

• DEFERRABLE constraints set to DEFERRED (INITIALLY DEFERRED or via SET CONSTRAINTS) are checked after each transaction.

Note the special treatment of UNIQUE / PRIMARY KEY constraints. Quoting the manual page for CREATE TABLE:

A constraint that is not deferrable will be checked immediately after every command.

While it states further down in the Compatibility section under Non-deferred uniqueness constraints:

When a UNIQUE or PRIMARY KEY constraint is not deferrable, PostgreSQL checks for uniqueness immediately whenever a row is inserted or modified. The SQL standard says that uniqueness should be enforced only at the end of the statement; this makes a difference when, for example, a single command updates multiple key values. To obtain standard-compliant behavior, declare the constraint as DEFERRABLE but not deferred (i.e., INITIALLY IMMEDIATE). Be aware that this can be significantly slower than immediate uniqueness checking.

Bold emphasis mine.

If you need any FOREIGN KEY constraints to reference the column(s), DEFERRABLE is not an option because (per documentation):

The referenced columns must be the columns of a non-deferrable unique or primary key constraint in the referenced table.